Differences in persistence between dogs and wolves in an unsolvable task in the absence of humans

Ethics statement

Special permission to use animals (wolves) in such cognitive studies is not required in Austria (Tierversuchsgesetz 2012). The ‘Tierversuchskommission am Bundesministerium für Wissenschaft und Forschung (Austria)’ allows research without special permissions regarding animals. We obtained ethical approval for this study from the ‘Ethik und Tierschutzcommission’ of the University of Veterinary Medicine (Protocol number ETK-07/08/2016).

Subjects

We tested 17 adult mixed-breed dogs (7 F, 10 M; mean age + SD = 4 + 1.6 years) (Canis lupus familiaris) and 12 adult Grey wolves (4 F, 8 M; mean age + SD = 6.3 + 1.7 years) (Canis lupus) raised and kept similarly in conspecific packs at the Wolf Science Centre, Austria (henceforth referred to as the ‘WSC’) (Table 1). The experiment lasted from October 2016 to February 2017. Subjects were hand-raised with conspecifics in peer groups by humans (dogs were raised separately from wolves and at different times). The animals had continuous access to humans who bottle-fed and later hand-fed them in the first 5 months of their life. During the first weeks of puppyhood, the animals were kept inside. They had free access to a 1,000 m² outdoor, ‘puppy’ enclosure from their second month on, and were moved to 2,000–8,000 m² ‘living’ enclosures at 5 months of age. The animals, as adults, live in these larger ‘home enclosures’. Packs are regularly moved from one home enclosure to another for logistic reasons (such as to make it easier to walk an animal on leash from its home enclosure to a test conducted indoors, or to a touristic event). All packs have resided in all home enclosures.

Every enclosure is equipped with bushes, trees, logs, shelters and permanent drinking water installations. While humans are not continuously present in living enclosures, all animals have regular social contact with humans as all animals voluntarily participate in cognitive and behavioural experiments, and/or training, and/or other social events at least once a day. Animals are rewarded with food for participating in these activities. This routine ensures that they are cooperative and attentive towards humans and allows weekly veterinary checks without sedation. All animals at the WSC are intact and males are vasectomized. Over the course of their lives, all animals at the WSC have participated in the same behavioural and cognitive experiments and have participated in the same training activities.

Apparatus

One object (henceforth referred to as the ‘ball’) was a perforated, hard plastic sphere 24 cm in diameter, weighing 1.5 kg (commercially available ‘Lion Feeder Ball’ from www.ottoenvironmental.com) (Fig. 1). The other object was a modified, perforated PVC sewage pipe (22 cm in diameter, 40 cm in length henceforth referred to as the ‘pipe’) (Fig. 2). Prior to the test, each object was baited with large chunks of strongly smelling sausage and meat out of sight of the subject.

Experimental setup

Before a test session began, we anchored one of the objects using a 30 cm long metal chain to a camping peg driven into the ground in the subjects’ home enclosure and marked a two m radius around it with a commercially available, bright red timber marking spray. This was done out of sight of the test subject. The peg was positioned such that any interactions the subject had with the object could be recorded from multiple angles without any visual obstructions. Two video cameras (recording at 1,920 × 1,080 pixels at 50 progressive frames per second) and one smartphone (Samsung Galaxy Note 2) were mounted on tripods outside the enclosures at three different angles. We used ‘IP Webcam,’ a freely available app developed by Pavel Khlebovich (http://ip-webcam.appspot.com/) to remotely monitor the trial, whilst staying out of sight of the subject during the entire procedure.

Subjects were tested in their home enclosure as they least expect a human to be present inside. Tests at the WSC are normally conducted in specific ‘testing enclosures’ and humans only visit the animals in the home enclosures in very specific contexts (i.e. pack visits, animal care, and short, training demonstrations during public guided tours for visitors and tourists). Subjects were in different home enclosures when they were tested with each object.

Procedure

We tested subjects individually between 7:00 and 10:00 a.m. One animal per pack was tested per session, and two to three sessions were conducted per week (never on consecutive days). To ensure high food motivation, the subjects were not fed the evening prior to testing. Before the test, we shifted the entire pack out of their home enclosure into an empty enclosure from where their home enclosure was out of sight. The test object was placed in the subjects’ home enclosure (see ‘Experimental Setup’) after which, the focal subject was led back into the home enclosure. We started the test session when the animal entered the two m-radius (see ‘Start’ in Table 2).

The subject was given five minutes to make first contact with the object. We defined ‘First Contact’ as the first time the subject touched or sniffed the object (in case of a sniff, when the nose was within five cm of the object). In case there was no ‘First Contact’ within 5 min, the test session was terminated. If the subject did not interact (i.e. ‘Sniff’ or ‘Manipulate’ the object—see Table 2 for definitions of all behaviours and behavioural states) with the apparatus at all for 5 min after ‘First Contact,’ the session was terminated. After the subject started interacting with the object, there was no limit on the duration it could continue to do so. Each time the subject stopped interacting with the object, we started a 5 min countdown. If the subject resumed interacting with the object before the countdown expired, we let the test session continue and reset the 5 min countdown timer. If the subject did not resume interacting with the object by the time the countdown expired, we terminated the test session. To simplify, if a subject started interacting with the object, it could continue doing so for an infinite duration and pause as many times as it liked, as long as each pause was shorter than 5 min; once it paused for more than 5 min, the test session ended.

After the session ended, we shifted the subject out of the home enclosure and retrieved the object. We carefully washed each object after each session to remove any possible odour cues left by the previously tested subject. Each subject was tested first with the ball and then, one and a half to three months later, re-tested with the pipe. Two wolves, Chitto and Tala, had to be tested with the pipe 6 months after their test with the ball due to the onset of the mating season. As we needed to keep our study comparable to a complementary study with free-ranging and pet dogs which were presented with only the ball (M. Lazzaroni et al., 2018, unpublished data), we were unable to counterbalance the presentation order of the two objects. We used each object only once per subject to avoid object-specific learning effects (e.g. to avoid subjects learning ‘food from inside this particular green sphere cannot be extracted’).

Behavioural coding

We recorded all tests on video and coded behaviours using Solomon Coder beta 100926 (a behaviour coding software developed by András Péter, Dept. of Ethology, Budapest; www.solomoncoder.com). We categorized manipulative behaviours based on the number of body parts involved and the nature of the behaviour. For instance, we differentiated between using paws to hold an object and to scratch vigorously at the object while the subject simultaneously gnawed at it. ‘Holding’ an object with the paws may have added stability which probably made ‘Biting’ more efficient, while ‘Scratching’ may not have added stability, but was perhaps a different strategy to extract the food within the object. The coded behaviours and their definitions are summarized in Table 2. See the Supplementary Video for an example of each behaviour. We defined ‘Persistence’ as the time (in seconds) a subject spent in the ‘Manipulating’ behavioural state. We defined ‘Contact Latency’ as the time (in seconds) a subject took from ‘Start’ to ‘First Contact.’ We defined ‘Motor Diversity’ as the number of unique ‘Manipulative Behaviours’ shown by a subject.

Analyses

We excluded one dog (Gombo) from the analyses for the pipe as he extracted some food from the object due to an apparatus malfunction (a piece of meat that we used had several long fibres that were too close to the holes in the apparatus, which allowed Gombo to easily grab them and pull a piece of meat out through one of the holes). We excluded one wolf (Una) from the latency analyses for the ball as her contact latency was an outlier (28 s; G = 5.09, U = 0.007, P < 0.001) (potentially because she was tested at the onset of the mating season). We excluded one dog (Nuru) from the analyses of the pipe as he was very unusually overly persistent with the pipe, making his manipulation duration an outlier (1,361 s, G = 3.10, U = 0.63, P = 0.008). We used Grubbs tests (Grubbs, 1950) with the package ‘outliers’ (version 0.14) (Komsta, 2006) in R version 3.4.3 (R Core Team, 2017) to confirm that these individuals were indeed outliers. See the Supplementary Material for how results changed when these latter two individuals were included in the analyses. All other subjects were included in the analyses (Ball: N = 11 wolves, 17 dogs, Pipe: N = 12 wolves, 15 dogs).

We used inter-class correlations (Shrout & Fleiss, 1979) implemented with the ‘psych’ package (version 1.7.8) (Revelle, 2017) in R version 3.4.3 (R Core Team, 2017) to calculate inter-observer reliability. A second coder coded 20% of the data and all variables achieved reliability coefficients between 0.89 and 0.99 between the two coders.

We first used an exploratory, principal component analysis (PCA) for each object to understand our data. Performing several univariate analyses may not have allowed us to understand the combined effect of all explanatory variables on our subjects’ task performance. As we were primarily interested in variables that have previously been shown to relate to problem-solving success, we included persistence, motor diversity, latency to contact, approach posture and likelihood of manipulation as explanatory variables. While we could have included several more variables (such as the frequencies of each manipulative behaviour), we chose to restrict the number of explanatory variables due to our relatively small dataset. We used the PCAmixdata package (version 3.1) (Chavent et al., 2014) in R (version 3.5.1) (R Core Team, 2018) which is designed for analysing multivariate data that is a mixture of continuous, discrete and categorical variables.

The PCAmixdata analysis algorithm classified subjects based on our explanatory variables which did not include ‘Species.’ The rationale behind leaving species out of the analysis was to allow the algorithm to classify subjects based purely on task performance and without any pre-existing bias. This way, if, for example there were distinct behavioural differences between the two species, it would result in data-point clusters composed entirely of dogs and entirely of wolves, and each cluster would have different values of one or more behavioural variables. Conversely, if there were no differences, it may still result in clusters with different variable values, but these clusters would be mixtures of dogs and wolves. We ran a separate multivariate analysis for each object as including data from both objects in one analysis made it difficult to meaningfully interpret the clusters’ structures. Separating the two objects allowed us to analyse whether subjects performed similarly with both objects. Additionally, we applied an orthogonal rotation procedure to each PCA to make interpretation easier. We used the ‘PCArot’ function, which uses a generalization of the varimax procedure for mixed data (Chavent, Kuentz-Simonet & Saracco, 2012). This procedure helps associate variables with a selected number of principal components (or dimensions) more clearly by providing either large (almost 1) or small (almost 0) loadings. While the variable loadings on each dimension (and hence the variance explained by each dimension) change after rotation, the total variance explained by the selected dimensions remains unchanged.

The PCA gave us useful insights into patterns in our data, but did not let us test whether there was a statistically significant difference in dogs’ and wolves’ performance when interacting with the two objects (we did not make any inferences based on the PCA results). Hence, we further analysed persistence, motor diversity and contact latency individually using generalized additive models for location, scale and shape (‘gamlss’ version 5.1-0) (Stasinopoulos & Rigby, 2007) in R version 3.5.1. We used the ‘gamlss.Dist’ package (version 5.0-6) to fit distributions to our data. We evaluated the distribution of each response variable and specified the best fitting distribution in the models. We evaluated model fits both by their generalized Akaike information criteria (AIC; Akaike, 1974) and by the distribution of the model residual quantile–quantile plots. This approach enabled us to analyse the data without major transformations (data transformations might have negatively affected our interpretations of the results (Feng et al., 2014; Lo & Andrews, 2015)).

To reduce the risk of our choice of distributions resulting in overfitting models to our data, we validated our models’ results by fitting identical models with other probable distributions, and compared models with different distributions but similar AIC values. Further, when our data fit multi-parametric variations of the same distribution equally well, we used the distribution with fewer parameters (e.g. ‘Persistence’ fit Weibull-1, Weibull-2 and Weibull-3 but we used Weibull-1, as this distribution is described with one parameter as against two or three). Results did not change between models, implying that they were robust against choice of distribution. For the sake of brevity, we have only reported results from models with the best fitting distributions here. Please see the Supplementary Material for the complete distribution selection, model selection and validation processes, outputs, and R scripts. To account for repeated measures, we included the individual as a random effect in all models that included ‘Object’ as a fixed effect. As our subjects ages varied, we included ‘Age’ as a fixed effect in all our models. This allowed us to account for any effects age may have on subjects’ task performance (Siwak, 2001). When interactions were not statistically significant, we ran a reduced model that included the same fixed and random effects but not the interaction term. We have reported the results from these, reduced models whenever interactions were not significant.

Based on the PCA’s results, we used a Fisher’s Exact Test in R version 3.5.1 to investigate whether dogs and wolves differed statistically in their likelihood to manipulate the objects. Our PCA suggested that wolves and dogs may differ in their persistence, but that this difference may be influenced by object type. To investigate this, we used a GAMLSS model to evaluate the effects of species, object type and a two-way interaction between them on the response variable persistence. To ensure model convergence, we added a miniscule constant (0.00001) to all persistence values. We fit this model with the Gamma distribution and validated it with the Box-Cox T Original, Weibull and Log-normal distribution. This process allowed us to achieve our first aim of testing our hypothesis about dog–wolf differences. We left motor diversity out of this analysis for two reasons: (1) our hypothesis pertained specifically to differences in persistence between dogs and wolves and (2) from our PCA (and from further analysis for our second aim), persistence and motor diversity appeared to be correlated; this collinearity may have negatively impacted our interpretation of model results (Graham, 2003).

For our second aim, we focussed on understanding the relationships between the correlates of problem-solving success within dogs and within wolves. We analysed data for both species separately by running separate GAMLSS models for dogs and wolves. The rationale behind this decision was that the only hypothesis we had pertaining to dog–wolf differences was about persistence and did not encompass other behavioural measures.

We ran two GAMLSS models with contact latency as the response variable. Our PCA suggested that contact latency may be related to object type, and that approach posture and persistence may influence contact latency differently in both objects. Hence, for dogs, we included object type, persistence, approach posture and two two-way interactions (object type by persistence and object type by approach posture) as explanatory variables. For dogs, we fit the model with the Inverse Gaussian distribution and validated it with the Inverse Gamma, Log-normal and Gamma distribution. For wolves, we fit the model with the Log-normal distribution and validated it with the Gamma, Weibull and Box-Cox Cole-Green distribution.

We ran two GAMLSS models with motor diversity as response variable. As the PCA suggested that persistence and motor diversity may be correlated, and because this correlation appeared slightly different between the two objects, we included persistence, object type and a two-way interaction between persistence and object type as explanatory variables. For dogs, we fit the model with the Zero Adjusted Poisson distribution and validated it with the Zero Inflated Poisson, Zero Adjusted Negative Binomial (Type I) and Zero Inflated Negative Binomial (Type I) distribution. For wolves, we fit the model with the Poisson distribution and validated it with the Zero Adjusted Poisson, Negative Binomial type I and Generalized Poisson Distribution.

Our last aim was to test subjects’ consistency in performance between the two tasks. As we had not restricted the duration a subject could manipulate both objects, and as contact latency could have varied due to the layout of the enclosure subjects were tested in, absolute persistence and latency values may not have been meaningfully comparable. Hence, we scaled these values from 0 to 1 in each task separately using the following formula for both variables: $V_s=\frac{V_i-\text{Min}\left(V_{\text{all}}\right)}{\text{Max}\left(V_{\text{all}}\right)-\text{Min}\left(V_{\text{all}}\right)}$ where V_s = scaled value (persistence or contact latency), V_i = individual’s unscaled value, Min/Max (V_all) = the minimum/maximum values for that object. We used a Spearman’s rank correlation on the scaled persistence and scaled contact latency data to test whether subjects were consistent in their persistence and contact latency between the two objects. We calculated a consistency score for persistence and contact latency by taking the absolute value of the difference between subjects’ scaled persistence scores (or scaled contract latency scores) for the ball and for the pipe. We used separate GAMLSS models to assess the effect of species on the consistency scores for persistence and contact latency. For persistence, we fit the model with the Generalized Beta Type 1 distribution and validated it with the Logit Normal distribution. For contact latency, we fit the model with the Simplex distribution and validated it with the Logit Normal and Beta Original distributions.

Multivariate analyses

The PCA for the ball produced five dimensions, the first three of which explained 83.28% of the variance in our data. Pre and post orthogonal rotation results are summarized in Table 3. The rotation significantly improved variable loadings on dimensions 1 and 3. Hence, we investigated these dimensions further. We found that dogs and wolves segregated into two near-distinct clusters along dimension 1, but not along dimension 3 (Fig. 3A). Persistence (0.82) and motor diversity (0.85) loaded very strongly on dimension 1 (Fig. 3B), suggesting that the segregation between dogs and wolves was likely due to differences in either persistence, motor diversity, or both, and that these two variables may be correlated. We found two distinct clusters along dimension 3, but each of these clusters were composed of both dogs and wolves. Approach posture loaded very strongly (0.98) on dimension 3 (as did contact latency, but to an almost negligible extent: 0.009). This suggested that there may be a very weak (if any) connection between contact latency and approach posture, and that neither of these variables were likely to be responsible for dog–wolf differences.

The PCA for the pipe also produced five dimensions, the first three of which explained 87.03% of the variance in our data. Pre and post orthogonal rotation results are summarized in Table 4. We investigated dimensions 1 and 2 further as the rotation significantly improved variable loadings on them. Unlike with the ball, wolves and dogs did not segregate into distinct clusters along either dimension (Fig. 4A). Persistence (0.68) and motor diversity (0.84) loaded strongly on dimension 1 (Fig. 4B), suggesting that these variables may be correlated. Like with the ball, approach posture and contact latency loaded strongly on different dimensions (Table 4; Fig. 4B). This supported results with the ball and suggested that there may not be a connection between contact latency and approach posture, and that neither variable contributed to dog–wolf differences.

Differences in persistence between wolves and dogs (GAMLSS)

Overall, 14 out of 17 dogs manipulated the ball and 10 out of 15 dogs manipulated the pipe. In contrast, all 11 wolves manipulated the ball and all 12 wolves manipulated the pipe. Wolves were significantly more likely to manipulate objects than dogs (Fisher’s Exact Test, Odds Ratio = 0.0, 95% CI [0.00–0.71], P = 0.015). Though the PCA suggested that persistence may have been affected by object type, the interaction between species and object was not significant (GAMLSS: t = −1.47, P = 0.15). Wolves were more persistent than dogs (GAMLSS: t = 3.73, P < 0.001) in their manipulation of the objects regardless of object-type (Fig. 5A). Neither subjects’ age (GAMLSS: t = 0.76, P = 0.45) nor object type (GAMLSS: t = 1.06, P = 0.29) affected persistence (Fig. 5B).

Relationship between correlates of problem-solving within dogs and within wolves

Contact latency decreased with persistence in both, dogs (GAMLSS: t = −4.35, P < 0.001) and wolves (GAMLSS: t = −3.42, P < 0.01). Neither the interaction between object type and persistence (GAMLSS; Dogs: t = 1.91, P = 0.07, Wolves: t = −0.96, P = 0.35) nor that between object type and approach posture significantly affected contact latency (GAMLSS; Dogs: t = −1.32, P = 0.20, Wolves: t = −1.61, P = 0.13). Neither object type (GAMLSS; Dogs: t = 1.44, P = 0.16, Wolves: t = −0.96, P = 0.35) nor approach posture (GAMLSS; Dogs: t = 0.43, P = 0.67, Wolves: t = −1.72, P = 0.10) significantly affected contact latency in either species. Contact latency decreased with age in dogs (GAMLSS: t = −2.85, P < 0.001) but not in wolves (GAMLSS: t = −0.04, P = 0.97).

Motor diversity increased with persistence in both, dogs (GAMLSS: t = 3.74, P < 0.001) and wolves (GAMLSS: t = 3.72, P = 0.001). The interaction between object type and persistence was not significant (GAMLSS; Dogs: t = −1.67, P = 0.11, Wolves: t = 1.62, P = 0.12). Neither object type (GAMLSS; Dogs: t = −1.74, P = 0.09, Wolves: t = −1.61, P = 0.12) nor age (GAMLSS; Dogs: t = −0.58, P = 0.57, Wolves: t = 1.20, P = 0.24) significantly affected motor diversity in either species.

Individual consistency

Both subjects’ persistence (Spearman’s ρ = 0.71, P < 0.001) and contact latency (Spearman’s ρ = 0.64, P < 0.001) across tasks were significantly correlated. Figure 6 shows individuals’ scaled persistence in both tasks. Overall, dogs were significantly more consistent both, in their persistence (GAMLSS: t = −2.31, P = 0.031) as well as in their contact latency (GAMLSS: t = −2.62, P = 0.02) than wolves.

For descriptive statistics of both groups’ performance in each task and for complete model information, see the Supplementary Material.